1 Introduction
In
Interpolative and Extrapolative Work—
The Place of Human Intelligence in the Age of Artificial Intelligence (hereafter
Interpolative and Extrapolative Work), Liyan Xu proposes that human intellectual work falls into two kinds, interpolative and extrapolative. The former operates within the boundaries of what is already known and fills in details, and the latter poses new questions and forges connections that did not previously exist. Current artificial intelligence (AI) is, at its core, an interpolation engine, so interpolative work should be delegated to AI, while human effort concentrates on the extrapolative frontier
[1]. This article broadly accepts this judgment, but argues that extrapolation contains a kind of activity that merits separate discussion, namely the reframing of problems. Problem reframing is still, in essence, a form of extrapolation. The word "beyond" in the title refers to the fact that the nature of this activity exceeds a presupposition implicit in the dichotomy between interpolation and extrapolation, namely that extrapolative work "will" gradually turn into interpolative work as data accumulate. In fact, within extrapolative work, problem reframing is precisely the part that is hardest to convert in this way.
Sometimes the posing of new questions and the forging of new connections between bodies of knowledge take place between two known fields, and once such a new connection is identified, it can be quickly summarized, learned, and replicated. At other times, however, raising a new question is closer to probing in the dark without any reference. One first proposes a direction that seems far-fetched. Most attempts end in failure, but a very few arrive at a position that no one had foreseen, and thereby change the state of an entire research field. This article calls the latter capacity imagination and, on this basis, puts forward two arguments. First, imagination is the part of extrapolation that is hardest for AI to absorb, because its value arises precisely outside existing evaluation standards. Second, imagination is not an innate gift that humans stably possess. After AI enters everyday work, the conditions under which people practice imagination may contract or be redistributed, depending on how workflows are arranged. Sustaining human imagination is therefore a value orientation that a society must actively adopt. The following sections attempt to argue for these two points.
2 Problem Reframing Within Extrapolation
One of the most direct manifestations of imagination is the reframing of the problem itself (sometimes what happens first is its deconstruction). Such reframing is not limited to major scientific revolutions, and instances of it can also be found in everyday life. For example, there is an online youth community whose interest centers on homophonic puns and linguistic deconstruction
[2], and it represents a certain "abstract culture" among contemporary youth
[3]. A common move in this culture is to re-segment a familiar word. An English counterpart is to read "important" as "import" plus "ant, " that is, an ant that has been imported.
① Such a reading unexpectedly does not extend outward along the original semantic relations, and instead changes the category to which the word belongs (cognitive linguistics calls this process frame-shifting
[4]). Repeating a ready-made pun is mere imitation, and playing to the crowd does not in itself constitute innovation. The value of this kind of "doing abstract" behavior lies in the moment of spark when someone first notices that "important" can be read as an imported ant, and that spark in itself carries a deep meaning of innovation. Viktor Shklovsky called this practice of interrupting habitual perception defamiliarization (остранение)
[5]. Wolfgang Pauli used the phrase "not even wrong" to describe propositions so absurdly wrong that they were not worth discussing in terms of right or wrong
②[6]. Following Pauli, a creative answer that deconstructs the question may be called "not even right" (in the sense of going beyond merely being right). The respondent temporarily steps outside the correctness standard of the original question and turns to examining whether that standard itself holds.
① The example in the Chinese version of this article is "明知山有虎, 那就不去明知山" ("There are tigers on Mingzhi Mountain, so don't go to Mingzhi Mountain"). It plays on the saying "明知山有虎, 偏向虎山行" ("knowing full well that there are tigers in the mountains, one still heads for them"), in which "明知" means "knowing full well." The joke, which has circulated widely in this community, re-segments "明知山" and reads it as the name of a mountain.
② A well-known example of "not even wrong" is answering "an obtuse angle" to the question "What is 1 + 1?"
Such questioning of existing definitions also appears in spatial practice. Every year, some temporary settlements are built in a desert during a specific period and dismantled after the event ends. These settlements have roads, blocks, public services, and operating rules, and accommodate tens of thousands of temporary residents for a short period. If such a settlement still cannot be considered a city, what other conditions must a city meet? Is permanence an indispensable condition of a city? This example raises the question of what a city is
[7]. Problem reframing in landscape architecture and spatial design often begins with a reexamination of established premises: reimagining the possibilities of an existing site and even going beyond the existing site to explore the possible forms of human settlement. Such questions rarely appear in the design briefs issued by a client, yet they may inspire discussions of the forms of human existence, and concern for the fate of humanity that matter more than the project itself.
Such problem reframing has already been discussed in the literature. Donald Schön called the process by which practitioners redefine the boundaries of a task "problem setting, " and pointed out that the situations practitioners face must often first be given a frame for action before they become problems that can be solved
[8]. Horst Rittel and Melvin Webber's account of "wicked problems" goes further. Planning problems often have no uncontested definition, and every formulation of a problem already contains a value position
[9]. Margaret Boden distinguished three types of creativity (combinational, exploratory, and transformational), of which transformational creativity changes the very rule space that generates solutions
[10]. These three bodies of work, from different disciplines, point to the same conclusion. Problem reframing is an independent intellectual operation, and it cannot be reduced to searching for better answers within a given problem (or generalizing outward within a certain range).
This conclusion bears directly on the dichotomy drawn in Interpolative and Extrapolative Work. Taking renderings as an example, Xu points out that extrapolative work that once occupied the frontier is gradually absorbed into the interpolable as data accumulate. This process holds for the connection-forging type of extrapolation distinguished at the beginning of this article. Once such connections are stably established, they enter the range of what AI can learn. Extrapolation of the problem-reframing type is different. It challenges the evaluation standards themselves, and data accumulated under the old standards cannot be used to understand a rejection of those standards. Problem reframing is therefore the hardest to bring into the scope of interpolation as data accumulate.
It should be noted that problem reframing is present throughout human work but is very sparse. Most human research and design also proceed within existing frameworks, and a large number of attempts at departure end in failure. Everyday language calls a successful moment of departure a "spark, " and the word itself contains a judgment about frequency. Sparse does not mean secondary. The directions of knowledge and design are often reset precisely by a few successful departures. Yet sparsity also has a methodological consequence for AI. Large language models are built on data volume, and the more frequently an event occurs, the more easily it is recognized as a pattern. Sparks of problem reframing sit in the low-frequency tail of any corpus, and each successful reframing changes the distribution of subsequent data, so earlier samples can hardly be used to predict the next one. Methods aimed at finding regularities have difficulty characterizing such events, and the evaluation of imagination cannot rely on average (or maximum likelihood) metrics. Its effect on the overall direction of human knowledge is concentrated in a few events that cannot be represented by probability. The natural next question, then, is where this capacity to depart from existing standards comes from.
3 Sources and Costs of Imagination
On the sources of imagination, the history of philosophy has offered some relatively conservative definitions. David Hume held that the materials of thought come from sensation and experience, and that all the mind can do is compound, transpose, augment, or diminish them
[11]. The performance of generative AI shows precisely why this definition is incomplete. The combinatorial capacity of contemporary large language models already far exceeds that of any individual. If imagination consisted only in combining old materials, models with massive combinatorial capacity should already possess imagination. Hume's definition therefore needs a supplement, and the supplement lies in selection based on aesthetics (or value judgment). Discussing mathematical creation, Henri Poincaré pointed out that the number of possible combinations is enormous, and that the real work lies in how the few "useful and beautiful" connections come to be selected, a selection that depends on the mathematician's aesthetic sensibility
[12]. Current models use statistical likelihood as the selection criterion in generation, and the aesthetic selection Poincaré described is not within their objective function. The sparsity described in Section 2 has a direct consequence here. Events that are low-frequency and that change the distribution sit exactly where likelihood-based selection has the greatest difficulty reaching (what geometry calls a cusp). Charles Peirce's notion of abduction further clarifies the nature of this selection. Faced with a surprising fact, the researcher first proposes an explanation that has not yet been verified, and invests work in it before the evidence arrives
[13]. A large model can generate any number of candidate explanations on instruction, but provisional trust in, and sustained investment in, one of these explanations cannot be supplied by the act of generation itself (although this can of course be done under human guidance). Taken together, imagination reconfigures the relations among materials and makes a committed selection among the results of that reconfiguration, and the materials themselves need not be new.
On this basis, imagination can be broken down into three stages, namely generating candidates that depart from existing standards, making a value-based selection among the candidates, and sustaining investment in the absence of support from existing evaluation standards. On instruction, AI can also generate candidates that are low-probability, unconventional, and even marked by frame-shifting. The first stage is therefore not unique to humans, and the difference between imagination as understood in this article and the generative capacity of AI is concentrated in the latter two stages, that is, selection and commitment. The example given in Section 2 can also be understood in this way. The reading of "important" as an imported ant could perhaps be enumerated by a model, but noticing that this reading is worth keeping, and that it carries a distinctive value in a social sense, is a human choice.
Such selection has observable behavioral features. In a longitudinal study of art students, Jacob Getzels and Mihaly Csikszentmihalyi found that students who explored their subject matter more openly and for longer in the early phase were later more likely to produce works judged to be original
[14]. High-level creators (note that the notion of "high-level" itself implies a value judgment) tend to settle the final form of the problem later. The counterpart in design education is that site investigations and sketches in the concept phase often go through several rounds of iteration before the designer settles on how to formulate the design problem, and the concept phase is precisely the part that design review criteria have the most difficulty covering. This feature runs opposite to the current training direction of large models. Large models are optimized to give complete and acceptable answers as quickly as possible, and deferring the settlement of the problem has no corresponding place in their training objectives. Simulated annealing can provide a limited analogy for the difference between these two ways of answering. The algorithm accepts temporarily worse moves (even random moves) with a certain probability, so that the search does not remain stuck in a local optimum
[15]. This is close to the off-the-point answers discussed above, in that both first leave the currently highest-rated position and enter a region that for the time being does not make sense. The analogy is limited because simulated annealing still obeys a given objective function, whereas human imagination may bring the objective function itself under doubt. The algorithm moves only within a given solution space, while humans can question the grounds on which that solution space is set.
③③ In theory, large models can also learn and train the hyperparameters on which they depend, that is, "lift" the problem up by one level. What this article seeks to point out, however, is the human capacity for "unlimited lifting."
Questioning the solution space itself leads imagination toward its mature form, and also toward its costs. René Descartes's method of doubt requires that all doubtful beliefs be temporarily suspended until an indubitable starting point is found
[16]. Willard Van Orman Quine further proposed that knowledge faces the test of experience as a whole, and that any part of it is in principle revisable
[17]. Karl Popper placed falsifiability at the core of the status of scientific theories
[18]. The three hold different positions, but all refuse to treat existing answers as the final boundary, and those who undertake such refusal must also bear the corresponding costs. Research may fail, a site design may not be accepted, and both the designer's professional reputation and the public interest may suffer. Existing AI can list counterexamples on instruction, but it has no reason to take on this risk of its own accord. Its generation of a sentence of objection differs in nature from researchers giving up a theory they have long believed. It can thus be seen that imagination depends on a set of fragile conditions. People must be willing to remain in uncertainty, must keep room for asking bad questions and drawing failed sketches, and must also be prepared to pay the price of failure.
This article cannot yet offer a complete explanation of the "ultimate" source of imagination. In terms of the three stages distinguished above, as to why a spark appears at a given moment, existing accounts can mostly offer only post hoc narratives. Discussing the Schema, Immanuel Kant pointed out that the Schema is always a product of the imagination (Einbildungskraft), and that the schematism of the understanding is a hidden art in the depths of the human soul (eine verborgene Kunst in den Tiefen der menschlichen Seele), whose true operations are hard to reveal
[19]. Popper likewise pointed out that the initial conception of a new idea does not fall within the scope of logical analysis, and that there is no logical method of arriving at new ideas
[18]. Section 2 described problem reframing as probing in the dark precisely because its starting point is hard to reduce to a function of past accumulation. What can be explained are the latter two stages. Judging whether a departure is worth persisting in and sustaining investment in depends on the practitioner's familiarity with the boundaries of existing frameworks, and this familiarity comes from the experience of working within those frameworks firsthand. Schön pointed out that a practitioner's capacity for problem setting rests on a "repertoire" accumulated across a large number of concrete cases. Facing a new situation, the practitioner sees it as a variant of some familiar case and thereby redefines the task
[8]. The aesthetic selection described by Poincaré likewise rests on long mathematical training
[12]. Only by revising a design again and again as it repeatedly fails to hold up can a practitioner tell which rules are real constraints and which are merely habits of the profession. What interpolative work supports, therefore, are the two stages of selection and commitment. The spark itself is not produced by accumulation, but without accumulation, people have no way of telling which spark is worth investing in. This is consistent with the claim in Section 2. Problem reframing is hard to bring into the scope of interpolation as data accumulate precisely because its starting point lies outside accumulation. What remains to be shown is that, as AI enters everyday work, this set of conditions is changing.
4 Contraction and Redistribution of Imagination
The change comes first from a training process that runs in the opposite direction. Humans train AI, and AI also trains humans. Current assistant-type AI models are trained with human feedback and tend to give answers that are relevant, helpful, clear, safe, and consistent with instructions
[20]. These requirements improve the usability of the tools, and they also form a dominant model of expression, with complete sentences, balanced structure, and an even-handed attitude. As users keep reading, accepting, and revising such outputs, over time they easily come to treat the model's "average" expression as normal expression, and to take high-probability AI lines of argument as the way thinking "has always been."
There is preliminary evidence for this concern. Anil Doshi and Oliver Hauser found that generative AI can raise the average ratings of individual short stories while reducing the overall diversity of a set of works
[21]. A cross-cultural experiment by Dhruv Agarwal et al. showed that AI writing suggestions make the writing of different participants more similar, and push the expression of Indian participants toward the American writing style with which the model is more familiar
[22]. Another experiment on poetry found that when people were placed in the position of revising finished AI output, their creativity was lower than in independent creation, and the gap disappeared only when the system placed people in the position of co-creators
[23]. These studies are not yet sufficient to prove that AI use necessarily harms imagination, but they already suffice to show that an improvement in the quality of works and an improvement in creative capacity are not necessarily the same thing.
In design disciplines, the same convergence may appear earlier, in problem setting. Highly reflective water surfaces, lush vegetation, and clean paving may make a single AI-generated image attractive enough, but with extensive use, such images gradually prescribe what kind of landscape is "worth" imagining. If designers start from generated images and then add site-based justification, they easily work backward from answers familiar to the model to the problem, and judgments that should have been formed through investigation, drawing, and repeated back-and-forth with the site are compressed into the screening of ready-made images. In this article we call this change "the interpolativization of humans." People increasingly occupy the supplementary positions reserved in model outputs, supplementing paragraphs, selecting schemes, and polishing images, while taking less and less part in deciding why a problem is posed the way it is. Section 2 pointed out that problem reframing is inherently sparse in intellectual activity. The other half of the reason can be added here, which is that people's opportunities to practice it are also shrinking. Telling which departures are worth persisting in depends on the experience of working within frameworks firsthand. When such work is handed over to AI as a whole, what diminishes is not only the opportunity to practice departure but also the conditions for forming this discernment.
The crux of the matter, therefore, is not whether design students or paper authors use AI, but the order of the workflow, that is, whether people call on tools after forming a problem, or start from the problems and answers the tools provide. Protecting imagination does not require rejecting AI, but it does require that people keep the parts of the workflow that cannot be outsourced. These include engaging with the site in person, posing the initial question themselves, giving explanations when values conflict, and keeping failed schemes. Once it is acknowledged that the way tools are used in turn shapes capacities, the division of labor between humans and AI can no longer be drawn solely on the basis of technical performance. This also means that the practice conditions discussed in Section 3 need not only contract after AI enters work. Humans have conditions that AI lacks. They perceive the physical world in person, experience spatial relationships on site, and form judgments through bodily experience, and these conditions do not disappear because AI enters the workflow. Once interpolative work such as drawing, retrieval, and text editing is taken over by AI, designers can in principle gain more time to stay on site and explore openly, that is, the conditions for longer exploration of problems in the sense of Getzels and Csikszentmihalyi
[14]. The two effects of contraction and release coexist, and AI use itself does not determine which prevails. What determines it is where the released time goes, and whether the workflow places people before or after the formation of the problem. The judgment of this article is therefore conditional. In workflows that start from answers provided by tools, the practice conditions contract, and the evidence of homogenization cited above describes precisely this kind of workflow. In workflows where people first form the problem and then call on tools, the practice conditions may be redistributed, or even expanded. Accordingly, the normative claim of this article is not limited to protecting a capacity that may be lost, and also includes cultivating this capacity through workflow and educational arrangements. In design education, site investigation, open-ended sketching, and the archiving of "failed" schemes all belong to such arrangements. This judgment should also distinguish the stage people are at. For practitioners who have already formed discernment, the time released by handing over repetitive work can be directed toward the workflow in which the problem is formed first. For students still in the learning stage, the same handover directly reduces accumulation itself, and this is precisely why design education needs to keep the parts in which basic work is done by hand.
5 Value Judgments Behind the Division of Labor
The proposal to delegate interpolative work to AI presupposes that interpolation is work in which judgment has already been made and only execution remains. This presupposition also deserves scrutiny. In model fitting, given a number of known points, one can draw many (even infinitely many) curves that pass through them (Fig. 1), and the known points do not tell us how the points "should" be connected. The concept of inductive bias in machine learning shows that a model must adopt some selection tendency in advance, beyond the data, or it cannot settle on one among the many functions that fit the data equally well
[24]. The no-free-lunch theorem shows that, detached from specific problems, there is no algorithm that is best for all optimization tasks
[25]. In design practice, under the same set of site conditions, there are still many schemes that satisfy technical codes, and whether to prioritize flood control or habitat is itself a ranking of values. Interpolation, then, also contains judgment, and delegating interpolation to AI as a whole also hands over part of the aesthetic judgments and trade-offs that ought to be made by humans.
Following this line of reasoning, determining whether a piece of work has already become interpolative is itself a value judgment, and the questions of who sets the criteria for this determination and on what basis they are adopted are easily passed over in discussion. Leaving technology to evolve on its own and letting the market choose appears to take no position, but in fact hands value judgments over to those who control models, data, computing power, and platforms. Langdon Winner pointed out long ago that technical artifacts are not neutral instruments detached from politics and institutions, and that they entrench particular power relations and modes of action
[26]. Further, for society, imagination is not only an individual quality but also a capacity at the organizational level of governments, firms, communities, and other types of organizations, and the conditions for imagination available to different parties are unequal. When the development of AI is led by those who control models, data, and computing power, this inequality is itself part of the value judgments behind the division of labor. Whose imagination is protected and who defines the futures worth imagining, on the one hand, and which work should be handed over to AI, on the other hand, are two sides of the same question. The workflows and institutions discussed in Section 4 are therefore not only arrangements of individual work habits, but also a distribution of the conditions for imagination among different social actors.
AI is still changing rapidly, and the boundary between interpolation and extrapolation is also moving. The composition of training data, the choice of objective functions and evaluation standards, and the direction of investment and regulation all influence where this boundary moves. Research on responsible innovation calls for anticipating impacts in advance during research and development, listening to the views of different groups, and adjusting direction in response to feedback
[27]. The notion of "sociotechnical imaginaries" in science and technology studies also reminds us that technological trajectories always carry visions of a shared future
[28]. Value guidance therefore cannot consist only of aligning with the past. Historical data preserve knowledge, but they also preserve biases and institutional inertia, and fitting the past more accurately is sometimes merely replicating the past more effectively. Iason Gabriel's analysis of value alignment distinguishes among people's immediate instructions, actual preferences, real interests, and the values they endorse upon reflection
[29]. The mainstream values of a society are not the average of existing preferences, and should also include criticism of preferences themselves, as well as the spirit of an era (Zeitgeist).
Finally, we return to the two arguments put forward at the beginning of this article. Imagination is the part of extrapolation that is hardest to absorb as data accumulate, and it is also the part most dependent on practice conditions, and whether these conditions contract or are redistributed after AI enters everyday work depends on where the workflow places people. Interpolative and Extrapolative Work provides a valuable starting point for discussing the division of labor between humans and AI. What this article seeks to add is that what truly needs to be protected and cultivated in this division of labor is not any specific type of job or skill, but the human capacity to depart from existing standards and explore in the dark, together with the workflows and institutions that protect and cultivate this capacity. How these workflows and institutions are distributed among different actors is likewise part of the question of the division of labor. If, amid convenience, people gradually lose the patience to remain in uncertainty, and are no longer willing to ask questions that may be wrong, or to draw a sketch that will most likely fail, then before being replaced by AI, people will already have trained themselves into more competent interpolators. Which work AI should take on is a question that must ultimately be answered by humans, and for which humans should be responsible.